In a significant move that underscores its ambitions beyond hardware, Xiaomi has officially unveiled MiMo-V2-Flash, a powerful new open-weight AI model designed to compete with industry leaders like DeepSeek, OpenAI, and Anthropic. This launch marks a pivotal step in Xiaomi’s “AGI roadmap,” signaling its intent to integrate advanced intelligence across its ecosystem of smartphones, tablets, and electric vehicles (EVs).
Technical Prowess and Efficiency MiMo-V2-Flash is built on a massive scale, boasting 309 billion parameters. To manage this size effectively, Xiaomi has utilized a Mixture-of-Experts (MoE) architecture. This framework allows the model to activate only specific sub-networks for different tasks, striking an optimal balance between high-level performance and computational efficiency. One of its standout features is its incredible speed, delivering up to 150 tokens per second, making it one of the fastest models in its class.
The model is also remarkably cost-effective. Xiaomi has priced the API at just $0.1 per million input tokens and $0.3 per million output tokens, significantly lower than many proprietary rivals. This aggressive pricing, combined with its open-weight nature—meaning developers can download and run it locally—positions it as an attractive alternative for the global developer community.
Performance and Real-World Application In benchmark testing, MiMo-V2-Flash has demonstrated elite capabilities. According to Xiaomi, the model matches the reasoning power of Moonshot AI’s Kimi K2 and DeepSeek V3.2. Remarkably, in coding evaluations like the SWE-Bench Verified, it achieved a score of 73.4%, rivaling Anthropic’s Claude 4.5 Sonnet despite being developed at a fraction of the cost.
Xiaomi’s internal teams, including experts like Luo Fuli (a former DeepSeek researcher), emphasized that the model utilizes a Hybrid Sliding Window Attention (SWA) mechanism. This technical choice helps the model handle long-context prompts without the usual lag or memory drain.
The Road Ahead The launch of MiMo-V2-Flash is more than just a software release; it is a foundational component of Xiaomi’s future strategy. President Lu Weibing noted that the deep integration of AI with the physical world is the “next frontier.” By making MiMo-V2-Flash available on platforms like Hugging Face and MiMo Studio, Xiaomi is fostering an open ecosystem that will likely see these AI “agents” eventually powering autonomous driving in Xiaomi EVs and providing proactive assistance on millions of handheld devices worldwide.
